Segmentation Fault w/ transform_jsonl()

[StephenTemp]

I believe I've installed m3-inference correctly, but running transform_jsonl() on a json lines file of tweets seems to fetch the first profile picture in the list and then terminate with a segmentation fault.

I believe the file is structured appropriately, in the format below:
{json object}\n
{json object}\n
...

Any idea what I might be running into?

[zijwang]

Does the example here work?

[StephenTemp]

It seems as though when I run the script everything works as intended; however, when I try to access via Python, I run into the seg faults. I've just tried running the test_transform_jsonl.py and got the same error.

[computermacgyver]

Thanks @StephenTemp for reporting this. Could you please share the version of Python and version of torch you have installed? Also what operating system are you on?

I would also be curious to know if the scripts/m3twitter.py file is working for you. E.g.,

python scripts/m3twitter.py --skip-cache --screen-name barackobama --auth scripts/auth.txt

To run that you'll need to create the auth.txt file in the same format as scripts/auth_example.txt .

I'd like to isolate whether it's the downloading and processing of images (which happens in both scripts/m3twitter.py and transform_jsonl) or if it is something specific to transform_jsonl.

[StephenTemp]

python version: 3.9.4
torch version: 1.9.0
operating system: macOS Big Sur (11.4 w/ M1)

It seems as though the scripts/m3twitter.py file is indeed working! Here's the output:

Also here's the specific error message:

[computermacgyver]

Thanks, @StephenTemp . I had a typo in Barack (missed the c) and as a result the example we tried didn't include an image. Could you re-run the test with

python scripts/m3twitter.py --skip-cache --screen-name barackobama --auth scripts/auth.txt

Or re-run with any profile that includes a picture. Sorry about that.

[StephenTemp]

No problem!

[computermacgyver]

Thanks, @StephenTemp . One more question is the version of PIL you have

python -c "import PIL; print(PIL.__version__)"

My best guess @zijwang is that this is somewhere in the download or resizing of images given the code runs on a profile with no image but fails on a profile with an image. Given it's a seg fault, it must be in compiled code, which leads me to think it's something with PIL or how we pass the image data to PIL in a BytesIO wrapper. It might be that we can work around this by loading the image from a file for resizing rather than passing it in memory.

m3inference/m3inference/preprocess.py

Lines 22 to 55 in bdaf727

	def download_resize_img(url, img_out_path, img_out_path_fullsize=None):
	# url=url.replace("_200x200","_400x400")
	try:
	img_data = urllib.request.urlopen(url)
	img_data = img_data.read()
	if img_out_path_fullsize != None:
	with open(img_out_path_fullsize, "wb") as fh:
	fh.write(img_data)
	except urllib.error.HTTPError as err:
	logger.warn("Error fetching profile image from Twitter. HTTP error code was {}.".format(err.code))
	return None
	except urllib.error.ContentTooShortError:
	logger.warn("Error fetching profile image from Twitter. The amount of data downloaded is less than the expected amount")
	return None
	except (urllib.error.URLError, ValueError) as err:
	logger.warn("Error fetching profile image from Twitter. {}".format(err))
	return None
	except:
	logger.warn("Unknown Error")
	return None
	return resize_img(BytesIO(img_data), img_out_path, force=True,url=url)
	def resize_img(img_path, img_out_path, filter=Image.BILINEAR, force=False,url=None):
	try:
	img = Image.open(img_path).convert("RGB")
	if img.size[0] + img.size[1] < 400 and not force:
	logger.info(f'{img_path} / {url} is too small. Skip.')
	return
	img = img.resize((224, 224), filter)
	img.save(img_out_path)
	except Exception as e:
	logger.warning(f'Error when resizing {img_path} / {url}\nThe error message is {e}\n')

I'll see if I can recreate on my OS (Linux) using the same versions.

[StephenTemp]

PIL: 8.2.0, thanks for your help!

[zijwang]

@StephenTemp I did a test run on 1) a valid screen name with a profile, 2) a valid screen name without a profile, and 3) a non-exist screen name with an Ubuntu machine and everything works just fine. Here is my pip list result:

Package            Version
------------------ ---------
certifi            2021.5.30
charset-normalizer 2.0.6
idna               3.2
m3inference        1.1.5
numpy              1.21.2
pandas             1.3.3
Pillow             8.3.2
pip                21.0.1
pycld2             0.41
python-dateutil    2.8.2
pytz               2021.1
rauth              0.7.3
requests           2.26.0
setuptools         58.0.4
six                1.16.0
torch              1.9.0
torchvision        0.10.0
tqdm               4.62.2
typing-extensions  3.10.0.2
urllib3            1.26.6
wheel              0.37.0

I also tried PIL 8.2.0 and things are still working fine.

One thing you may try is to add more log prints to m3inference/m3inference/preprocess.py and see which exact line throws the segfault. This will provide us more insight on what the root cause is.

[zijwang]

@StephenTemp -- checking in to see whether this issue has been resolved :)

[computermacgyver]

operating system: macOS Big Sur (11.4 w/ M1)

Looks like this could be the same issue as #26

It might be specially related to the arm64/M1 chipset, which would explain why we couldn't reproduce it

[StephenTemp]

Apologies for the wait; it's been a cluttered semester! Yes, unfortunately I couldn't work around the transform_json() function but was able to run the inference itself. It seems that the failure occurs in scripts/m3twitter.py on line
pprint.pprint(m3Twitter.infer_screen_name(args.screen_name, skip_cache=args.skip_cache))